Visualizing Feature Maps in Deep Neural Networks using DeepResolve - A Genomics Case Study, Ge Liu, David Gifford

Although many powerful visualization tools have been developed to interpret neural network decisions in input space, methods to interpret feature map space remain limited. Most existing tools examine a network’s response to a specific input sample and thus are locally faithful to that sample. We introduce DeepResolve, a gradient ascent based method that visualizes intermediate layer feature maps in an input independent manner. We examine DeepResolve’s capability to 1) discover network linear and non-linear combinatorial logic and summarize overall knowledge of a class, 2) reveal key features for a target class, 3) assess a network’s activeness in pattern learning and network’s vulnerability in feature space, and 4) analyze multi-task class similarity at high resolution. We demonstrate the value of DeepResolve on synthetic and experimental genomic datasets, and DeepResolve reveals biologically interesting observations from the experimental data.